High-precision hyperdimensional computing enabled by in-memory computing using high-polarization ferroelectric Hf_0.5Zr_0.5O₂ capacitors

Abstract

We demonstrate high-precision hyperdimensional computing using an in-memory computing (IMC) architecture based on ferroelectric Hf0.5Zr0.5O2 (HZO) capacitors. By exploiting the high polarization charge density of CMOS back-end-compatible HZO, we achieved 32 well-separated and linearly programmable intermediate states in 10-nm-thick capacitors making them suitable as capacitive IMC elements. In recent times, capacitive IMC emerged as a promising energy- and latency-efficient route for data-intensive computing tasks. However, compute-in-memory elements require non-volatile, reproducible, and multi-bit operation. In this work, we show that through optimized device fabrication without vacuum break between oxide and nitride depositions and tailored thermal engineering, the HZO capacitors can exhibit high remanent polarization (2Pr = 75 µC/cm²). Structural studies highlight a high orthorhombic phase fraction and clean HZO/TiN interface. The intermediate polarization states exhibit controllable, linear, and reproducible capacitance modulation via voltage-driven polarization switching, enabling reliable multi-bit device operation and non-destructive readout. Leveraging these 5-bit ferroelectric capacitors, it is possible to store 15-bit numerical values using only three capacitors to implement high-precision capacitive IMC in a hyperdimensional computing task, achieving improved inference accuracy of 92.3% and 2.3x reduced areal footprint compared to binary encoding. These results highlight the importance of advanced materials engineering to achieve high bit-precision and state linearity in ferroelectric capacitors for scalable capacitive in-memory computing. (Figure presented.)